MountainScape semantic segmentation of historical and repeat images
Date
2025
Authors
Mahindrakar, Aniket
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Semantic segmentation of ultra-high resolution images is challenging due to high memory and computation requirements. Current approaches to this problem involve cropping the ultra-high resolution image into small patches for individual processing in order to provide local context, or under-sampling the images to provide global context, or following a combination of both which gives rise to global-local refinement pipelines. In this thesis, we present the MountainScape Segmentation Dataset (MS2D) which comprises high-resolution historic (grayscale) manually segmented images of Canadian mountain landscapes captured from 1861 to 1958 and their corresponding modern (colour) repeat images. Additionally, we analyze the characteristics of the dataset, define evaluation criteria, and provide a baseline to serve as a reference benchmark for automated land cover classification using the Python Landscape Classification Tool (PyLC), an existing software tool. The main contribution of this thesis is the experimental exploration of various deep learning architectures to address the tiling artifacts and spatial context loss faced by PyLC in its tile-based processing of ultra-high-resolution images, alongside a comprehensive investigation using a larger dataset than that employed in the original PyLC study to solve this tiling problem.
Description
Keywords
Machine learning, Landcover classification, Oblique images, Mountain Legacy Project, Remote Sensing